Improved reliability of fiber orientation estimation and graph theoretical analysis of structural brain networks with diffusion MRI

Date: 2 June 2017

Venue: UAntwerp, Campus Middelheim, A.143 - Middelheimlaan 1 - 2020 Antwerpen (route: UAntwerpen, Campus Middelheim)

Time: 4:00 PM

PhD candidate: Timo Roine

Principal investigator: J. Sijbers, A. Leemans & B. Jeurissen

Short description: PhD defence Timo Roine - Faculty of Science, Department of Physics


Diffusion magnetic resonance imaging (MRI) has the unique ability to non-invasively measure brain microstructure and structural connectivity in vivo. Until recently, the accuracy of the reconstructed neural pathways has been limited due to the inability of diffusion tensor imaging (DTI) to detect crossing pathways, present in the majority of white matter. Recent high angular resolution diffusion imaging (HARDI) methods overcome this obstacle and thus, enable more reliable measures for structural brain connectivity.

Constrained spherical deconvolution (CSD) is one of these methods, and can reliably estimate the full fiber orientation distribution by deconvolving the diffusion MRI data with a response function describing a single coherently oriented fiber population. However, non-white matter partial volume effects can significantly decrease the precision of the identified fiber orientations and increase the number of false positives, most prominently at the interface between the white and gray matter of the brain. These effects dramatically affect the reconstruction of neural tracts with diffusion MRI.

The whole-brain reconstruction of the neural pathways, the connectome, forms a large and complex network connecting different brain structures to each other. Both global and local network properties can be extracted from the connectome by using graph theoretical analysis. However, the reproducibility of the network properties and their intercorrelations in CSD-based structural connectomes have not been thoroughly investigated. This may cause researchers to select irreproducible parameters for their analyses and thus, decrease the probability of finding true effects.

This dissertation aims to improve the reliability of reconstructing neural fibers in the brain by accounting for the non-white matter partial volume effects in CSD. Moreover, it provides insight into the reproducibility and intercorrelation in graph theoretical analysis of the structural brain networks.

First, we analyzed the fiber orientations estimated with CSD under non-white matter partial volume effects (PVEs), and estimated that these PVEs affect 35–50 % of white matter voxels. Through simulations, we found that the non-white matter PVEs decreased the precision of the estimated fiber orientations and increased the number of identified false fiber orientations.

To account for the non-white matter PVEs in CSD, we developed a new method, informed CSD. We showed with simulations that by using anatomical information to incorporate the PVEs into the response function used in CSD, the precision of the estimated fiber orientations could be improved and the number of false positives could be significantly reduced. The simulation results were confirmed by using real diffusion MRI data and residual bootstrapping.

To assess the reproducibility of graph theoretical analysis in structural brain networks, we analyzed data from 19 healthy subjects. We performed residual bootstrapping to simulate nine repeated acquisitions, which resulted in a total of ten datasets per subject. We calculated intraclass correlation coefficients and coefficients of variation for each subject and various network properties, weights and reconstruction parameters. The results showed that a reconstruction density of approximately one million streamlines is required for excellent reproducibility, and that binarization of the networks should be done with significantly higher threshold values than only one streamline. Based on the results, we provide guidelines for the reproducible analysis of the structural brain networks.